The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
1140 
1.5°E 2°E 2.5“E 
Figure 1: Geographic layout of evaluated CARTOSAT-1 
scenes and ICC reference data used for verification. 
3.2 RPC correction 
Matching and thus GCP collection between CARTOSAT-1 and 
ETM+ scenes is hindered by the large time and resolution 
differences. Most scenes of the ETM+ mosaic have been 
captured between 1999 and 2001, resulting in large differences 
between CARTOSAT-1 and ETM+ scenes. Due to the large 
differences in appearance, relatively loose thresholds on 
correlation (0.7) and bidirectional matching coordinate shift 
(0.5 pixels in the ETM+ image) have been used during GCP 
collection. A sufficient number of well distributed GCP is 
found for all 9 scenes. An iterative outlier removal procedure is 
applied during the affine RPC correction estimation. Between 
856 and 93 GCP are used for RPC correction for RPC 
correction. The image space residuals of the GCP are quite 
large, with standard deviation between 1.7 and 2 pixels, mostly 
due to the large resolution difference between CARTOSAT-1 
and ETM+. No systematic error is visible in the residuals. 
The C-SAP demonstrated that subpixel residuals can be 
achieved when a few high quality and well distributed GCP are 
available (Lehner et al., 2007). From the above results, it is 
expected that the Landsat ETM+ mosaic is not a suitable base 
for deriving GCP with the accuracy required for CARTSAT-1. 
The DSM based RPC refinement described in section 2.2.1 is 
thus used to further reduce the error. To estimate the true 
accuracy of the two RPC correction approaches, 68 checkpoints 
have been measured in the ICC orthoimages and the stereo 
partner Cat-A. The height of each checkpoint is derived from 
the ICC DTM. These measurements have been automatically 
transformed into Cat-A/F tie points via least squares matching. 
6 window sizes from 17 to 27 have been used in LSM in order 
to get statistical values for the accuracy. Forward intersection of 
these stereo tie points results in object space positions. The 
lateral and height differences between stereo points and 
checkpoints are given in Table 2. It is obvious that the 
correction based on GCP derived from ETM+ and SRTM leads 
to a high shift in location and height. Considering the 15 m 
resolution of ETM+, a mean difference of 12.5 m is still a good 
result and indicates subpixel accuracy of the ETM+ Geocover 
mosaic in the studied area. After aligning the stereo points to 
SRTM and re-estimation of the affine RPC correction, the 
lateral displacement reduces to 3.5 m. This is a very good result, 
especially when considering the 90 m grid spacing of the 
SRTM. Figure 2 shows the lateral shifts of all checkpoints. 
Figure 2: Lateral error of Cat scene for the two affine RPC 
correction methods, measured using independent 
checkpoints. Red, thick arrows: ETM+ and SRTM 
GCP. Blue, thin arrows: Align to SRTM. The arrow 
lengths are scaled by a factor 200. 
RPC 
correction 
reference 
Lateral difference (m) 
Mean o 
Height difference 
(m) 
Mean a 
ETM+, 
SRTM 
12.51 
3.25 
1.20 
2.40 
Align to 
SRTM 
3.48 
1.10 
0.30 
1.47 
Table 2: Accuracy of the two RPC correction procedures, 
measured using well distribured, independent 
checkpoints. 
3.3 Forward intersection and outlier removal 
Forward intersection of the mass stereo tie points performs well; 
no points are discarded with a relatively strict threshold of 0.5 
pixels on the image space residuals. A few gross blunders with 
mismatches along the epipoles remain and are rejected due to 
their deviation from the SRTM DSM. 
Table 3 shows the number of accepted points and their mean 
height difference to the ICC reference DTM. When comparing 
the differences with the checkpoint evaluation in section 3.2, 
the larger height errors and standard deviations are noticeable. 
This is caused by the suboptimal conditions for image matching, 
such as the low sun angle and the mostly mountainous terrain 
with vegetation. Especially in the Montseny mountain range 
located in the upper right of the block, very large black 
shadows with a diameter of several km can be found, leading to 
large interpolation facets and resulting in a rather coarse DSM 
with large facets. Most of scene 117/209 is covered by the 
ocean and the city of Barcelona with large build up areas. When 
comparing the generated surface model with the bare earth 
DTM provided by ICC, a negative height difference, as well as 
a larger standard deviation is expected in such areas. The 
negative mean height difference observed for all scenes is a 
good sign and shows that the CARTOSAT DSM is located 
above the ICC DTM.